@genai-fi/nanogpt 0.11.0 → 0.12.1

package/dist/layers/PositionEmbedding.js

@@ -1,9 +1,9 @@
-import { t as c, al as u, b as i } from "../index-D0RBWjq8.js";
+import { t as c, al as u, b as i } from "../index-Duu1Lvvv.js";
 import f from "./BaseLayer.js";
-import { E as g, D as h } from "../random_width-UGQn4OWb.js";
-import { r as l } from "../exports_initializers-eS9QJ6ut.js";
-import { m as b } from "../mod-PrOKlFxH.js";
-import { r as w } from "../range-CuGvVN2c.js";
+import { E as g, D as h } from "../random_width-BN4wGJaW.js";
+import { r as l } from "../exports_initializers-1KWPiStI.js";
+import { m as b } from "../mod-B4AUd1Np.js";
+import { r as w } from "../range-DKmP1-OQ.js";
 function E(t) {
   return new h(t);
 }

package/dist/layers/RMSNorm.js

@@ -1,8 +1,8 @@
-import { t as s } from "../index-D0RBWjq8.js";
+import { t as s } from "../index-Duu1Lvvv.js";
 import e from "./BaseLayer.js";
 import { normRMS as a } from "../ops/normRMS.js";
-import { v as i } from "../variable-DQ9yYgEU.js";
-import { o as m } from "../ones-BX_wEgzB.js";
+import { v as i } from "../variable-DPt_Iuog.js";
+import { o as m } from "../ones-CBI1AQjb.js";
 class l extends e {
   GAMMA;
   constructor(r, t = "", o) {

package/dist/layers/RoPECache.js

@@ -1,7 +1,7 @@
-import { l as a, b as n, p as c, t as p, k as r } from "../index-D0RBWjq8.js";
-import { r as l } from "../zeros_like-BAz3iKru.js";
-import { c as f, s as m } from "../unsorted_segment_sum-LAbmE9G4.js";
-import { r as h } from "../range-CuGvVN2c.js";
+import { j as a, b as n, p as c, t as p, k as r } from "../index-Duu1Lvvv.js";
+import { r as l } from "../zeros_like-De4n1C3m.js";
+import { c as f, s as m } from "../unsorted_segment_sum-BljxHhCY.js";
+import { r as h } from "../range-DKmP1-OQ.js";
 class D {
   constructor(s) {
     this.config = s;

package/dist/layers/TiedEmbedding.js

@@ -1,13 +1,13 @@
-import "../random_width-UGQn4OWb.js";
-import "../index-D0RBWjq8.js";
-import { r as s } from "../exports_initializers-eS9QJ6ut.js";
+import "../random_width-BN4wGJaW.js";
+import "../index-Duu1Lvvv.js";
+import { r as s } from "../exports_initializers-1KWPiStI.js";
 import a from "./BaseLayer.js";
 import { dot16 as o } from "../ops/dot16.js";
 import { isPackedTensor as r } from "../utilities/packed.js";
-import { p as m } from "../pack16-k4jq6aMX.js";
+import { p as m } from "../pack16-atD0eYRm.js";
 import { transpose16 as d } from "../ops/transpose16.js";
-import { v as p } from "../variable-DQ9yYgEU.js";
-import { g as h } from "../gather-BcO5UQNJ.js";
+import { v as p } from "../variable-DPt_Iuog.js";
+import { g as h } from "../gather-BSULDalH.js";
 class g extends a {
   vocabSize;
   embedDim;

package/dist/layers/TransformerBlock.js

@@ -2,7 +2,7 @@ import p from "./CausalSelfAttention.js";
 import m from "./MLP.js";
 import d from "./RMSNorm.js";
 import h from "./BaseLayer.js";
-import { k as n, t as u } from "../index-D0RBWjq8.js";
+import { k as n, t as u } from "../index-Duu1Lvvv.js";
 import { add16 as l } from "../ops/add16.js";
 class g extends h {
   ln1;

package/dist/loader/loadTransformers.js

@@ -1,7 +1,7 @@
 import l from "../tokeniser/CharTokeniser.js";
 import c from "../tokeniser/bpe.js";
 import { load_safetensors as b } from "../utilities/safetensors.js";
-import { aa as y } from "../index-D0RBWjq8.js";
+import { a9 as y } from "../index-Duu1Lvvv.js";
 import { dummyPassAsync as u } from "../utilities/dummy.js";
 import _ from "../models/factory.js";
 async function L(e, a, r, t) {

package/dist/loader/oldZipLoad.js

@@ -1,22 +1,22 @@
-import { aa as y } from "../index-D0RBWjq8.js";
-import "../random_width-UGQn4OWb.js";
-import "../zeros_like-BAz3iKru.js";
+import { a9 as y } from "../index-Duu1Lvvv.js";
+import "../random_width-BN4wGJaW.js";
+import "../zeros_like-De4n1C3m.js";
 import "../Generator.js";
 import "../index-Cp39cXWe.js";
-import "../dataset-DcjWqUVQ.js";
+import "../dataset-CVIJu7Xa.js";
 import "../ops/cpu/attentionMask.js";
 import "../ops/webgl/attentionMask.js";
 import "../ops/grads/attentionMask.js";
 import "../ops/cpu/rope.js";
 import "../ops/webgl/rope.js";
-import "../rope-BmZmp9uP.js";
+import "../rope-DJ7Y7c-u.js";
 import "../ops/cpu/appendCache.js";
 import "../ops/webgl/appendCache.js";
 import "../ops/grads/softmax16.js";
-import "../matMul16-cDxwemKj.js";
+import "../matMul16-xswmhSuF.js";
 import "../ops/webgl/matMul16.js";
 import "../ops/cpu/matMul16.js";
-import "../pack16-k4jq6aMX.js";
+import "../pack16-atD0eYRm.js";
 import "../ops/transpose16.js";
 import "../ops/reshape16.js";
 import "../ops/cpu/qkv.js";
@@ -42,26 +42,26 @@ import "../ops/cpu/adamMoments.js";
 import "../ops/webgl/adamMoments.js";
 import "../papaparse.min-C0cScC2i.js";
 import "../ops/cpu/matMulGelu.js";
-import "../matMulGelu-B2s_80-H.js";
+import "../matMulGelu-BpvgnYG8.js";
 import "../ops/grads/matMulGelu.js";
 import "../ops/cpu/gelu.js";
 import "../ops/webgl/gelu.js";
-import "../gelu-DqTbCx5x.js";
+import "../gelu-BK1k-n1i.js";
 import "../ops/webgl/log.js";
 import "../checks/normRMS.js";
 import "../checks/normRMSGrad.js";
 import { importWeights as u } from "../utilities/weights.js";
 async function Mt(r) {
-  const e = /* @__PURE__ */ new Map(), a = await r.file("manifest.json")?.async("string");
-  if (!a)
+  const e = /* @__PURE__ */ new Map(), p = await r.file("manifest.json")?.async("string");
+  if (!p)
     throw new Error("Manifest file not found in the zip archive");
-  const p = JSON.parse(a);
-  for (const [t, o] of Object.entries(p.weightSpec))
+  const s = JSON.parse(p);
+  for (const [t, o] of Object.entries(s.weightSpec))
     e.set(t, { spec: o, data: new Float32Array() });
-  const s = await r.file("tokeniser.json")?.async("string");
-  if (!s)
+  const a = await r.file("tokeniser.json")?.async("string");
+  if (!a)
     throw new Error("Tokeniser file not found in the zip archive");
-  const i = JSON.parse(s), f = (i.type ?? "char") === "char" ? new h(i.vocab) : new k(i.vocab, i.merges), c = /* @__PURE__ */ new Map();
+  const i = JSON.parse(a), f = (i.type ?? "char") === "char" ? new h(i.vocab) : new k(i.vocab, i.merges), c = /* @__PURE__ */ new Map();
   for (const t of Object.keys(r.files))
     if (t.endsWith(".bin")) {
       const o = t.replace(".bin", ""), l = await r.file(t).async("arraybuffer"), w = new Float32Array(l), n = e.get(o) || { spec: [], data: new Float32Array() };
@@ -70,7 +70,7 @@ async function Mt(r) {
       c.set(o, d);
     }
   y();
-  const m = b(p.config);
+  const m = b(s.config);
   return await g(m), m.loadWeights(c), { model: m, tokeniser: f };
 }
 export {

package/dist/{log_sum_exp-VLZgbFAH.js → log_sum_exp-CVqLsVLl.js}

@@ -1,7 +1,7 @@
-import { q as r, u as x, E as p, am as E, an as $, ao as d, af as h, c as S, n as K } from "./index-D0RBWjq8.js";
-import { e as _ } from "./axis_util-DofAuy0p.js";
-import { r as m } from "./reshape-CkjKPPqB.js";
-import { s as T } from "./sum-BdplSvq_.js";
+import { o as r, q as x, E as p, am as E, an as $, ao as d, af as h, c as S, l as K } from "./index-Duu1Lvvv.js";
+import { e as _ } from "./axis_util-DGqbT-FX.js";
+import { r as m } from "./reshape-BI0yzp1T.js";
+import { s as T } from "./sum-BPUfDB2X.js";
 function b(s, o = null, n = !1) {
   const a = { x: x(s, "x", "max") }, e = { reductionIndices: o, keepDims: n };
   return p.runKernel(E, a, e);

package/dist/main.d.ts

@@ -1,3 +1,6 @@
+import { default as PretrainingTask } from './training/tasks/PretrainingTask';
+import { default as StartSentenceTask } from './training/tasks/StartSentenceTask';
+import { default as ConversationTask } from './training/tasks/ConversationTask';
 import { pack16 } from './ops/pack16';
 import { unpack16 } from './ops/unpack16';
 import { default as CausalSelfAttention } from './layers/CausalSelfAttention';
@@ -19,6 +22,12 @@ export type { TrainingProgress, TrainingLogEntry } from './training/Trainer';
 export type { GPTConfig } from './models/config';
 export { estimateParameterCount, estimateMemoryUsage, estimateTrainingMemoryUsage, estimateResources, validateConfig, } from './utilities/parameters';
 export { default as topP } from './utilities/topP';
+export { Task, tokensFromTasks } from './training/tasks/Task';
+export declare const tasks: {
+    PretrainingTask: typeof PretrainingTask;
+    StartSentenceTask: typeof StartSentenceTask;
+    ConversationTask: typeof ConversationTask;
+};
 declare const ops: {
     pack16: typeof pack16;
     unpack16: typeof unpack16;

package/dist/main.js

@@ -1,18 +1,22 @@
-import "./index-D0RBWjq8.js";
-import "./random_width-UGQn4OWb.js";
-import "./zeros_like-BAz3iKru.js";
-import { default as oo } from "./Generator.js";
+import "./index-Duu1Lvvv.js";
+import "./random_width-BN4wGJaW.js";
+import "./zeros_like-De4n1C3m.js";
+import { default as io } from "./Generator.js";
 import "./index-Cp39cXWe.js";
-import "./dataset-DcjWqUVQ.js";
-import { default as to } from "./models/NanoGPTV1.js";
-import { default as mo } from "./TeachableLLM.js";
-import { default as io } from "./tokeniser/CharTokeniser.js";
-import { default as so } from "./tokeniser/bpe.js";
-import { default as lo } from "./utilities/waitForModel.js";
-import { default as xo } from "./data/textLoader.js";
-import { default as co } from "./models/model.js";
-import { estimateMemoryUsage as ko, estimateParameterCount as go, estimateResources as Mo, estimateTrainingMemoryUsage as Po, validateConfig as Co } from "./utilities/parameters.js";
-import { default as ho } from "./utilities/topP.js";
+import "./dataset-CVIJu7Xa.js";
+import { default as fo } from "./models/NanoGPTV1.js";
+import { default as lo } from "./TeachableLLM.js";
+import { default as co } from "./tokeniser/CharTokeniser.js";
+import { default as ko } from "./tokeniser/bpe.js";
+import { default as go } from "./utilities/waitForModel.js";
+import { default as Co } from "./data/textLoader.js";
+import { default as Eo } from "./models/model.js";
+import { estimateMemoryUsage as Bo, estimateParameterCount as bo, estimateResources as yo, estimateTrainingMemoryUsage as Lo, validateConfig as So } from "./utilities/parameters.js";
+import { default as Ao } from "./utilities/topP.js";
+import { Task as Go, tokensFromTasks as Ro } from "./training/tasks/Task.js";
+import o from "./training/tasks/PretrainingTask.js";
+import r from "./training/tasks/StartSentenceTask.js";
+import t from "./training/tasks/ConversationTask.js";
 import "./ops/cpu/scatterSub.js";
 import "./ops/webgl/scatterSub.js";
 import "./ops/cpu/gatherSub.js";
@@ -25,15 +29,15 @@ import "./ops/webgl/qkv.js";
 import "./ops/grads/qkv.js";
 import "./ops/cpu/rope.js";
 import "./ops/webgl/rope.js";
-import "./rope-BmZmp9uP.js";
+import "./rope-DJ7Y7c-u.js";
 import "./ops/cpu/appendCache.js";
 import "./ops/webgl/appendCache.js";
 import "./ops/cpu/matMulGelu.js";
-import "./matMulGelu-B2s_80-H.js";
+import "./matMulGelu-BpvgnYG8.js";
 import "./ops/grads/matMulGelu.js";
 import "./ops/cpu/gelu.js";
 import "./ops/webgl/gelu.js";
-import "./gelu-DqTbCx5x.js";
+import "./gelu-BK1k-n1i.js";
 import "./ops/cpu/normRMS.js";
 import "./ops/webgl/normRMS.js";
 import "./ops/grads/normRMS.js";
@@ -42,51 +46,58 @@ import "./ops/cpu/adamMoments.js";
 import "./ops/webgl/adamMoments.js";
 import "./ops/cpu/adamAdjust.js";
 import "./ops/webgl/adamAdjust.js";
-import { u as o, p as r } from "./pack16-k4jq6aMX.js";
+import { u as e, p as m } from "./pack16-atD0eYRm.js";
 import "./ops/grads/softmax16.js";
-import "./matMul16-cDxwemKj.js";
+import "./matMul16-xswmhSuF.js";
 import "./ops/webgl/matMul16.js";
 import "./ops/cpu/matMul16.js";
 import "./ops/transpose16.js";
-import { selectBackend as bo } from "./backend.js";
-import { default as Lo } from "./utilities/performance.js";
-import t from "./layers/CausalSelfAttention.js";
-import e from "./layers/MLP.js";
-import m from "./layers/TransformerBlock.js";
-import p from "./layers/RoPECache.js";
-import { default as Go } from "./training/AdamExt.js";
-import { default as Uo } from "./checks/index.js";
-import { sentenceEmbeddings as wo, sentenceEmbeddingsTensor as Do } from "./utilities/sentences.js";
-const Y = {
-  pack16: r,
-  unpack16: o
-}, Z = {
-  CausalSelfAttention: t,
-  MLP: e,
-  TransformerBlock: m,
-  RoPECache: p
+import { selectBackend as wo } from "./backend.js";
+import { default as No } from "./utilities/performance.js";
+import p from "./layers/CausalSelfAttention.js";
+import a from "./layers/MLP.js";
+import i from "./layers/TransformerBlock.js";
+import s from "./layers/RoPECache.js";
+import { default as qo } from "./training/AdamExt.js";
+import { default as Ho } from "./checks/index.js";
+import { sentenceEmbeddings as Jo, sentenceEmbeddingsTensor as Ko } from "./utilities/sentences.js";
+const to = {
+  PretrainingTask: o,
+  StartSentenceTask: r,
+  ConversationTask: t
+}, eo = {
+  pack16: m,
+  unpack16: e
+}, mo = {
+  CausalSelfAttention: p,
+  MLP: a,
+  TransformerBlock: i,
+  RoPECache: s
 };
 export {
-  Go as AdamExt,
-  so as BPETokeniser,
-  io as CharTokeniser,
-  oo as Generator,
-  co as Model,
-  to as NanoGPT,
-  mo as TeachableLLM,
-  Uo as checks,
-  ko as estimateMemoryUsage,
-  go as estimateParameterCount,
-  Mo as estimateResources,
-  Po as estimateTrainingMemoryUsage,
-  Z as layers,
-  xo as loadTextData,
-  Y as ops,
-  Lo as performanceTest,
-  bo as selectBackend,
-  wo as sentenceEmbeddings,
-  Do as sentenceEmbeddingsTensor,
-  ho as topP,
-  Co as validateConfig,
-  lo as waitForModel
+  qo as AdamExt,
+  ko as BPETokeniser,
+  co as CharTokeniser,
+  io as Generator,
+  Eo as Model,
+  fo as NanoGPT,
+  Go as Task,
+  lo as TeachableLLM,
+  Ho as checks,
+  Bo as estimateMemoryUsage,
+  bo as estimateParameterCount,
+  yo as estimateResources,
+  Lo as estimateTrainingMemoryUsage,
+  mo as layers,
+  Co as loadTextData,
+  eo as ops,
+  No as performanceTest,
+  wo as selectBackend,
+  Jo as sentenceEmbeddings,
+  Ko as sentenceEmbeddingsTensor,
+  to as tasks,
+  Ro as tokensFromTasks,
+  Ao as topP,
+  So as validateConfig,
+  go as waitForModel
 };

package/dist/{matMul16-cDxwemKj.js → matMul16-xswmhSuF.js}

@@ -1,9 +1,9 @@
-import { j as y, e as h } from "./index-D0RBWjq8.js";
+import { i as y, e as h } from "./index-Duu1Lvvv.js";
 import "./ops/webgl/matMul16.js";
 import "./ops/cpu/matMul16.js";
 import { isPackedTensor as g } from "./utilities/packed.js";
-import { p as v } from "./pack16-k4jq6aMX.js";
-import { d as k } from "./gelu-DqTbCx5x.js";
+import { p as v } from "./pack16-atD0eYRm.js";
+import { d as k } from "./gelu-BK1k-n1i.js";
 import { transpose16 as S } from "./ops/transpose16.js";
 import { reshape16 as w } from "./ops/reshape16.js";
 const G = {
@@ -54,7 +54,7 @@ function l(r, o, n = !1, s = !1, t = {}) {
   const e = g(r), f = g(o), i = e || f, a = !i || e ? r : v(r), p = !i || f ? o : v(o), c = h().runKernel("MatMul16", { A: a, B: p }, { transposeA: n, transposeB: s, ...t });
   return i && !e && a.dispose(), i && !f && p.dispose(), c;
 }
-function F(r, o, n, s = !1, t = !1) {
+function K(r, o, n, s = !1, t = !1) {
   return l(r, o, s, t, { scale: n });
 }
 function B(r, o, n, s = !1, t = !1) {
@@ -63,14 +63,14 @@ function B(r, o, n, s = !1, t = !1) {
 function M(r, o, n, s = !1, t = !1) {
   return l(r, o, s, t, { scaleB: n });
 }
-function K(r, o, n = !1, s = !1) {
+function N(r, o, n = !1, s = !1) {
   return l(r, o, n, s, { activation: "gelu" });
 }
 export {
   G as a,
   l as b,
-  K as c,
+  N as c,
   B as d,
   M as e,
-  F as m
+  K as m
 };

package/dist/{matMulGelu-B2s_80-H.js → matMulGelu-BpvgnYG8.js}

@@ -1,12 +1,12 @@
-import { h as C, t as R, e as I, V as G, $ as L, y as F, ak as U } from "./index-D0RBWjq8.js";
-import { r as M } from "./Reshape-Ct266DEk.js";
-import { u as H } from "./gpgpu_math-CJcbnKPC.js";
-import { m as B } from "./mat_mul-DxpNTCRz.js";
+import { f as C, t as R, e as I, U as G, _ as L, x as U, ak as F } from "./index-Duu1Lvvv.js";
+import { r as M } from "./Reshape-Ck29jQSY.js";
+import { u as H } from "./gpgpu_math-BJSTk_mW.js";
+import { m as B } from "./mat_mul-Bn2BDpT4.js";
 class W {
   constructor(e, s, a, n = !1, o = !1, r = !1, i = null, u = !1, l = !1) {
     this.variableNames = ["matrixA", "matrixB"], this.packedInputs = !0, this.packedOutput = !0, this.outputShape = a, this.enableShapeUniforms = H(this.outputShape.length);
-    const p = n ? e[1] : e[2], h = Math.ceil(p / 2), d = n ? "i * 2, rc.y" : "rc.y, i * 2", x = o ? "rc.z, i * 2" : "i * 2, rc.z", b = n ? ["a.xxyy", "a.zzww"] : ["a.xxzz", "a.yyww"], m = o ? ["b.xzxz", "b.ywyw"] : ["b.xyxy", "b.zwzw"];
-    let c = "", $ = "";
+    const p = n ? e[1] : e[2], h = Math.ceil(p / 2), d = n ? "i * 2, rc.y" : "rc.y, i * 2", b = o ? "rc.z, i * 2" : "i * 2, rc.z", x = n ? ["a.xxyy", "a.zzww"] : ["a.xxzz", "a.yyww"], m = o ? ["b.xzxz", "b.ywyw"] : ["b.xyxy", "b.zwzw"];
+    let c = "", g = "";
     i && (u ? c = `vec4 activation(vec4 a) {
           vec4 b = getPreluActivationWeightsAtOutCoords();
           ${i}
@@ -15,8 +15,8 @@ class W {
           ${i}
         }` : c = `vec4 activation(vec4 x) {
           ${i}
-        }`, $ = "result = activation(result);");
-    const g = r ? "result += getBiasAtOutCoords();" : "";
+        }`, g = "result = activation(result);");
+    const $ = r ? "result += getBiasAtOutCoords();" : "";
     r && this.variableNames.push("bias"), u && this.variableNames.push("preluActivationWeights"), l && this.variableNames.push("leakyreluAlpha");
     let f = "rc.x", v = "rc.x";
     e[0] < s[0] ? f = `imod(rc.x, ${e[0]})` : s[0] < e[0] && (v = `imod(rc.x, ${s[0]})`), this.userCode = `
@@ -30,12 +30,12 @@ class W {
         int batchB = ${v};
         for (int i = 0; i < ${h}; i++) {
           vec4 a = getMatrixA(batchA, ${d});
-          vec4 b = getMatrixB(batchB, ${x});
+          vec4 b = getMatrixB(batchB, ${b});
 
           // These swizzled products need to be separately added.
           // See: https://github.com/tensorflow/tfjs/issues/1735
-          result += (${b[0]} * ${m[0]});
-          result += (${b[1]} * ${m[1]});
+          result += (${x[0]} * ${m[0]});
+          result += (${x[1]} * ${m[1]});
         }
         return result;
       }
@@ -44,10 +44,10 @@ class W {
         ivec3 rc = getOutputCoords();
         vec4 result = dot2x2ARowBCol(rc);
 
-        ${g}
-
         ${$}
 
+        ${g}
+
         setOutput(result);
       }
     `;
@@ -90,30 +90,30 @@ function O({
   activationSnippet: o,
   multiplier: r
 }) {
-  const i = t.shape.length, u = e.shape.length, l = s ? t.shape[i - 2] : t.shape[i - 1], p = a ? e.shape[u - 1] : e.shape[u - 2], h = s ? t.shape[i - 1] : t.shape[i - 2], d = a ? e.shape[u - 2] : e.shape[u - 1], x = t.shape.slice(0, -2), b = e.shape.slice(0, -2), m = G(x), c = G(b), g = L(t.shape.slice(0, -2), e.shape.slice(0, -2)).concat([h, d]);
-  F(
+  const i = t.shape.length, u = e.shape.length, l = s ? t.shape[i - 2] : t.shape[i - 1], p = a ? e.shape[u - 1] : e.shape[u - 2], h = s ? t.shape[i - 1] : t.shape[i - 2], d = a ? e.shape[u - 2] : e.shape[u - 1], b = t.shape.slice(0, -2), x = e.shape.slice(0, -2), m = G(b), c = G(x), $ = L(t.shape.slice(0, -2), e.shape.slice(0, -2)).concat([h, d]);
+  U(
     l === p,
     () => `Error in matMul: inner shapes (${l}) and (${p}) of Tensors with shapes ${t.shape} and ${e.shape} and transposeA=${s} and transposeB=${a} must match.`
   );
-  const f = s ? [m, l, h] : [m, h, l], v = a ? [c, d, p] : [c, p, d], A = M({ inputs: { x: t }, backend: n, attrs: { shape: f } }), y = M({ inputs: { x: e }, backend: n, attrs: { shape: v } }), k = [A, y], E = Math.max(m, c), N = o, T = U(t.dtype, e.dtype), _ = new W(
+  const f = s ? [m, l, h] : [m, h, l], v = a ? [c, d, p] : [c, p, d], A = M({ inputs: { x: t }, backend: n, attrs: { shape: f } }), y = M({ inputs: { x: e }, backend: n, attrs: { shape: v } }), k = [A, y], _ = Math.max(m, c), E = o, N = F(t.dtype, e.dtype), T = new W(
     f,
     v,
-    [E, h, d],
+    [_, h, d],
     s,
     a,
     !1,
-    N,
+    E,
     !!r,
     !1
   ), D = [A, y];
   r && D.push(r);
-  const z = n.runWebGLProgram(_, D, T), K = M({ inputs: { x: z }, backend: n, attrs: { shape: g } });
+  const z = n.runWebGLProgram(T, D, N), K = M({ inputs: { x: z }, backend: n, attrs: { shape: $ } });
   k.push(z);
   for (const P of k)
     n.disposeIntermediateTensorInfo(P);
   return K;
 }
-function V(t) {
+function J(t) {
   const { inputs: e, backend: s } = t, { x: a, kernel: n } = e;
   if (a === void 0 || n === void 0)
     throw new Error("BatchMatMul requires two input tensors.");
@@ -126,13 +126,13 @@ function V(t) {
     activationSnippet: j
   });
 }
-const J = {
+const Q = {
   kernelName: "MatMulGelu",
   backendName: "webgl",
-  kernelFunc: V
+  kernelFunc: J
 };
-C(J);
-function Q(t) {
+C(Q);
+function V(t) {
   const { dy: e, x: s, kernel: a } = t.inputs, n = t.backend;
   return R(() => {
     const o = I().makeTensorFromTensorInfo(
@@ -152,12 +152,12 @@ function Q(t) {
 const X = {
   kernelName: "MatMulGeluGrad",
   backendName: "webgl",
-  kernelFunc: Q
+  kernelFunc: V
 };
 C(X);
 export {
   W as M,
   se as a,
   O as b,
-  V as c
+  J as c
 };

package/dist/mat_mul-Bn2BDpT4.js

@@ -0,0 +1,11 @@
+import { o as m, q as s, B as c, E as M, D as p } from "./index-Duu1Lvvv.js";
+function f(e, o, n = !1, l = !1) {
+  let a = s(e, "a", "matMul"), t = s(o, "b", "matMul");
+  [a, t] = c(a, t);
+  const r = { a, b: t }, u = { transposeA: n, transposeB: l };
+  return M.runKernel(p, r, u);
+}
+const i = /* @__PURE__ */ m({ matMul_: f });
+export {
+  i as m
+};

package/dist/{mod-PrOKlFxH.js → mod-B4AUd1Np.js}

@@ -1,4 +1,4 @@
-import { q as r, u as s, D as m, E as c, M as d } from "./index-D0RBWjq8.js";
+import { o as r, q as s, B as m, E as c, M as d } from "./index-Duu1Lvvv.js";
 function p(t, e) {
   let o = s(t, "a", "mod"), a = s(e, "b", "mod");
   [o, a] = m(o, a);

package/dist/models/NanoGPTV1.js

@@ -3,11 +3,11 @@ import b from "../layers/TransformerBlock.js";
 import k from "../layers/TiedEmbedding.js";
 import w from "../layers/RoPECache.js";
 import E from "../layers/RMSNorm.js";
-import { t as l, k as u } from "../index-D0RBWjq8.js";
+import { t as l, k as u } from "../index-Duu1Lvvv.js";
 import C from "./model.js";
 import P from "../layers/PositionEmbedding.js";
 import { packingSupported as _ } from "../utilities/packed.js";
-import { p as y, u as M } from "../pack16-k4jq6aMX.js";
+import { p as y, u as M } from "../pack16-atD0eYRm.js";
 class I extends C {
   wte;
   // Token embeddings

package/dist/models/model.js

@@ -1,23 +1,23 @@
 import m from "../layers/BaseLayer.js";
-import "../index-D0RBWjq8.js";
-import "../random_width-UGQn4OWb.js";
-import "../zeros_like-BAz3iKru.js";
+import "../index-Duu1Lvvv.js";
+import "../random_width-BN4wGJaW.js";
+import "../zeros_like-De4n1C3m.js";
 import "../Generator.js";
 import "../index-Cp39cXWe.js";
-import "../dataset-DcjWqUVQ.js";
+import "../dataset-CVIJu7Xa.js";
 import "../ops/cpu/attentionMask.js";
 import "../ops/webgl/attentionMask.js";
 import "../ops/grads/attentionMask.js";
 import "../ops/cpu/rope.js";
 import "../ops/webgl/rope.js";
-import "../rope-BmZmp9uP.js";
+import "../rope-DJ7Y7c-u.js";
 import "../ops/cpu/appendCache.js";
 import "../ops/webgl/appendCache.js";
 import "../ops/grads/softmax16.js";
-import "../matMul16-cDxwemKj.js";
+import "../matMul16-xswmhSuF.js";
 import "../ops/webgl/matMul16.js";
 import "../ops/cpu/matMul16.js";
-import "../pack16-k4jq6aMX.js";
+import "../pack16-atD0eYRm.js";
 import "../ops/transpose16.js";
 import "../ops/reshape16.js";
 import "../ops/cpu/qkv.js";
@@ -40,11 +40,11 @@ import "../ops/webgl/scatterSub.js";
 import "../ops/cpu/gatherSub.js";
 import "../ops/webgl/gatherSub.js";
 import "../ops/cpu/matMulGelu.js";
-import "../matMulGelu-B2s_80-H.js";
+import "../matMulGelu-BpvgnYG8.js";
 import "../ops/grads/matMulGelu.js";
 import "../ops/cpu/gelu.js";
 import "../ops/webgl/gelu.js";
-import "../gelu-DqTbCx5x.js";
+import "../gelu-BK1k-n1i.js";
 import "../ops/webgl/log.js";
 import "../checks/normRMS.js";
 import "../checks/normRMSGrad.js";

package/dist/{ones-BX_wEgzB.js → ones-CBI1AQjb.js}

@@ -1,6 +1,6 @@
-import { w as n, W as t, V as m, E as i } from "./index-D0RBWjq8.js";
-import { c as f } from "./complex-DClmWqJt.js";
-import { z as c } from "./zeros-DeiE2zTa.js";
+import { u as n, V as t, U as m, E as i } from "./index-Duu1Lvvv.js";
+import { c as f } from "./complex-Cyg-eQeZ.js";
+import { z as c } from "./zeros-5YROwwUH.js";
 function l(o, r = "float32") {
   if (n(o), r === "complex64") {
     const s = l(o, "float32"), a = c(o, "float32");

package/dist/ops/adamAdjust.js

@@ -1,4 +1,4 @@
-import { e as i } from "../index-D0RBWjq8.js";
+import { e as i } from "../index-Duu1Lvvv.js";
 import "./cpu/adamAdjust.js";
 import "./webgl/adamAdjust.js";
 function p(r, t, e, n, m, o) {

package/dist/ops/adamMoments.js

@@ -1,4 +1,4 @@
-import { e as t } from "../index-D0RBWjq8.js";
+import { e as t } from "../index-Duu1Lvvv.js";
 import "./cpu/adamMoments.js";
 import "./webgl/adamMoments.js";
 function s(e, n, r, m, o) {

package/dist/ops/add16.js

@@ -1,4 +1,4 @@
-import { n as t, e as o } from "../index-D0RBWjq8.js";
+import { l as t, e as o } from "../index-Duu1Lvvv.js";
 import { isPackedTensor as n } from "../utilities/packed.js";
 import "./grads/add16.js";
 function m(r, e) {